7929 results about "Computation complexity" patented technology

The invention relates to the technical field of image detection, in particular to a PCBA board defect detection method and system based on image processing, and the method comprises the following steps: carrying out the meshing calculation of a gray scale deviation after a gray scale image is subjected to Gaussian filtering denoising, generating change rate data, carrying out the statistics of a frequency number, constructing a histogram, combining with an Otsu algorithm, and generating a candidate mask; extracting pixels based on a mask, calculating a gradient modulus, screening edge candidate points, carrying out gradient direction connection and morphological processing to generate a complete edge structure, expanding a connected domain through a region growing algorithm, aligning the connected domain with a template contour, and outputting defect coordinates. According to the method, the defect identification sensitivity is improved through combination of gray level image gridding processing and dynamic threshold calculation, a candidate mask is generated through grid gray level change rate statistics and an Otsu algorithm to avoid over-segmentation missing detection, and the contour precision is improved through combination of gradient modulus difference screening and morphological closed operation optimization. The region growing algorithm and template dynamic alignment reduce deformation misjudgment, and staged dimension reduction and feature enhancement reduce calculation complexity and solve resource waste.

The invention relates to the technical field of wind power prediction. The invention provides a wind power prediction method and system. The method comprises the following steps: acquiring multi-dimensional meteorological time series data, three-dimensional elevation data and unit operation data of a target wind power plant; constructing a spatial-temporal feature fusion network, extracting time sequence dynamic features, and performing weighted fusion on the spatial correlation features and the time sequence dynamic features to obtain a fusion feature vector; establishing a hybrid prediction model, and taking the fusion feature vector as input to obtain a wind power initial prediction result; introducing a terrain correction factor, constructing a turbulence intensity compensation function, and performing micro-terrain disturbance correction on the wind power initial prediction result; and outputting a final power prediction curve and a confidence interval. The problems that in an existing wind power prediction method, a physical model is insufficient in complex terrain microclimate modeling precision, high in calculation complexity and difficult to meet the real-time requirement, a statistical learning method is limited in high-dimensional nonlinear time sequence feature expression capacity, and prediction errors are remarkably increased under the abnormal working condition are solved.

The invention discloses a large model reasoning efficiency dynamic optimization and hardware sensing compression method. The method comprises the following five steps: S1, generating an input complexity signal representing calculation complexity; s2, synchronously monitoring a hardware resource index of the operation platform, and generating a hardware state signal reflecting a real-time load; s3, inputting the input complexity signal and the hardware state signal into a dynamic strategy selector, and generating a compression control signal through a pre-trained decision model; s4, according to the compression control signal, dynamic reconfiguration operation is executed on the large model weight and the activation value of the current reasoning task; and S5, performing reasoning calculation by using the reconfigured large model, and feeding back a hardware resource index to the step S2 in real time in the calculation process to form a closed-loop optimization link. According to the large model reasoning efficiency dynamic optimization and hardware perception compression method, the problems of low resource utilization rate, delay fluctuation and energy efficiency imbalance caused by a static compression method in dynamic input and heterogeneous hardware environments can be solved.

The invention belongs to the field of natural language processing, and provides a task complexity driven graph semantic multi-agent collaborative decision-making method and system.The task complexity driven graph semantic multi-agent collaborative decision-making method comprises the steps that a task text is obtained and subjected to semantic coding to obtain a task semantic vector, evaluation is conducted based on the task semantic vector to obtain a complexity vector, and a task complexity score of the complexity vector is calculated; the task semantic vector and the complexity vector are fused to obtain a task representation vector, an agent capability relation graph is constructed, the participation probability of each agent node is obtained according to the task representation vector and the agent capability relation graph, and a dynamic agent combination scheme is formed; and performing task decomposition according to the agent combination scheme, constructing a sub-task dependency graph, scheduling the execution sequence of the sub-tasks through topological sorting, realizing cooperative execution of the agents, and generating a task result. According to the method, precise matching and efficient cooperation of the agent combination are realized, and the capability of processing complex tasks and the resource utilization efficiency of the multi-agent system are remarkably improved.

The invention provides an adaptive multi-modal feature fusion well logging interpretation method based on multi-task joint learning, and the model comprehensively employs a graph convolution network, a multi-head attention mechanism, a gating mechanism and an adaptive loss weight adjustment algorithm, and effectively integrates multi-well group well logging data and geological structure information. And the complex spatial relationship and geological similarity between the wells are fully captured. The model can process multi-source data in a unified manner, extracts high-level feature representation containing spatio-temporal correlation, avoids repeated feature learning, reduces calculation complexity, and ensures the consistency and complementarity of prediction task features at the same time. The self-adaptive loss weight adjustment mechanism dynamically optimizes weight distribution according to the loss change of each task, realizes balance optimization among different tasks, and comprehensively improves the prediction precision of each task. The physical constraint regularization module ensures that the prediction result not only has statistical accuracy, but also conforms to the actual geological physical law, so that the reasonability and reliability of the prediction result are improved.

The invention provides a methane leakage anomaly detection system and method based on dual-light fusion, and belongs to the field of industrial safety detection. According to the system, a double-light-path holder type laser detector is arranged in a monitoring area, and methane concentration and image data are synchronously acquired through a 1653nm near-infrared laser module and a visible light image module. Time features (mean values, trends and the like) are extracted from the concentration data through time sequence analysis, and spatial features (edges, textures and the like) are extracted from the image data through deep learning; the system fuses time and space features to generate a comprehensive feature vector, reduces the calculation complexity through a dimension reduction technology, and retains core features at the same time. And constructing an anomaly detection model based on an unsupervised learning algorithm, identifying anomaly features in real time, and triggering an alarm. The system realizes real-time and accurate detection of methane leakage, and enhances the reliability and efficiency of industrial safety.

The invention discloses a segmentation method for rail defect detection based on improved YOLOv10 and SETR, and the method comprises the steps: S1, collecting image data, carrying out the enhancement processing of the image data, carrying out the segmentation processing, and making a data set; s2, an improved YOLOV10 algorithm model is built and trained; s3, using the improved YOLOV10 model weight to detect image data to be detected, judging whether defects exist or not, if the defects are detected, storing detection frame information of the defects, and outputting a target detection frame and confidence; and S4, transmitting the detection frame information as prompt information into the SETR model, segmenting the target by adopting the SETR model to obtain a segmentation result, and outputting a mask pattern. The improved YOLOv10 algorithm and the SETR algorithm are adopted to detect the rail defects, the calculation complexity and the efficient detection speed are kept, the detection precision is improved, and the robustness of the model is improved.

The invention discloses a hyperspectral image classification method based on proxy attention and a VIT architecture, and belongs to the field of image classification, and the method comprises the following steps: preprocessing hyperspectral image data, and dividing the hyperspectral image data into a training set and a test set; constructing a hyperspectral image classification model based on proxy attention and a VIT architecture; training the hyperspectral image classification model by adopting the training set; and inputting the hyperspectral data in the test set into the trained hyperspectral image classification model, reasoning to obtain a prediction category, and evaluating the obtained prediction category through an evaluation index. According to the invention, the agent attention encoder based on agent attention is constructed, hierarchical feature extraction is carried out through the agent attention encoder of L-layer cascade, layer-by-layer fusion and enhancement of space-spectrum information in an initial feature coding matrix are realized, the calculation complexity is significantly reduced while the strong representation capability is maintained, and the method has the advantages of high efficiency, high resolution and the like. And good balance between the calculation efficiency and the feature expression capability is realized.

The invention relates to the technical field of artificial intelligence and natural language processing, in particular to a key value cache compression and sparse attention calculation method and system for large language model reasoning, and the method comprises the steps: an offline calibration stage; the online reasoning stage comprises the following steps: a pre-filling step; an autoregression generation step: for each newly generated lexical element, projecting a current query vector Q and a key vector K in a key cache to a low-dimensional space to obtain Q'and K '; calculating an approximate attention score based on Q'and K ', and selecting an index I of the first k most relevant lexical elements which are ranked from high to low; and calculating an accurate attention score based on Q and K [I], and calculating with the value vector V [I] to obtain the output of the current lexical element. According to the scheme, the memory and calculation bottleneck of large model reasoning in a scene of long text sequence input are solved, and the method has the advantages of reducing video memory occupation and calculation complexity at the same time.

The invention discloses a bearing residual life prediction method based on a multi-teacher element weight knowledge distillation network, and the method comprises the following steps: (1) carrying out the normalization, noise reduction and segmentation of a multi-dimensional vibration time sequence signal collected by an original sensor, and dividing the signal into a training set and a verification set; (2) inputting the processed training set data into a multi-teacher element weight knowledge distillation network, training teacher models, and fixing parameters of the three teacher models after training is completed; (3) performing knowledge distillation training on the student model by using an adaptive meta-weight strategy gradient learning algorithm; and (4) the student model after distillation training is used for bearing residual life prediction. According to the method, the number of network model parameters is small, the calculation complexity is low, the prediction precision is high, and the method can be actually deployed on edge equipment.

The invention relates to the technical field of instrument multi-task optimization, in particular to an intelligent instrument multi-task real-time optimization method and system based on dynamic resource scheduling. The optimization method comprises the following steps: acquiring a target item of each task in real time through a sensor array, constructing a multi-dimensional feature vector, dividing each task into task categories by using a fuzzy clustering algorithm, and presetting an initial priority for the task categories for multi-task feature parameter acquisition and classification modeling. According to the method, the multi-dimensional feature vectors including the task urgency degree, the calculation complexity and the data interaction frequency are constructed, and the fuzzy clustering algorithm of the task dependency constraint is introduced, so that the task categories are accurately divided, the cross-category interaction overhead of the dependency task is effectively reduced, and the compatibility of a scheduling strategy is improved from the source.

The invention discloses a capacitor appearance intelligent defect detection method based on YOLO-NAS and Transform fusion, and belongs to the field of artificial intelligence and industrial automation, a lightweight backbone network is combined with MobileViT, a feature extraction architecture is optimized through NAS, and the global feature expression ability is improved; a Transform-FPN (Fabry-Perot Networking) is introduced to realize multi-scale feature fusion; instructing student model learning by utilizing high-order features of the teacher model through feature distillation; optimizing small target and irregular defect detection precision by adopting an improved positioning loss function; and the Soft-NMS and the DIOU-NMS are combined, and the calculation complexity is reduced through model pruning and hybrid quantization, so that the real-time detection deployment of the embedded equipment is realized. According to the capacitor appearance intelligent defect detection method based on YOLO-NAS and Transform fusion, the problems of missing detection of small targets, difficulty in distinguishing multiple defects and high model calculation complexity are solved by combining a lightweight backbone network, a Transform module, multi-scale feature fusion and an improved loss function.

The invention relates to an aviation turbine blade service damage monitoring method based on digital twinning, belongs to the technical field of aero-engines, and solves the problems that an existing detection algorithm is high in calculation complexity and low in efficiency and precision. Determining a to-be-monitored turbine blade physical entity, and collecting turbine blade environment information; a digital twin is constructed on the basis of a physical entity, the turbine blade three-dimensional model construction module constructs a turbine blade three-dimensional model according to the physical entity, and the turbine blade operation environment simulation module performs environment simulation on the model according to real-time load, rotation speed and environment temperature data of a turbine blade. And the turbine blade defect evolution prediction module identifies blade surface defects of the model at each moment by using an intelligent defect detection algorithm to obtain the category, position information and size information of the defects, and performs evolution simulation and danger coefficient analysis. According to the invention, real-time on-line, high-precision and high-efficiency detection of the surface defects of the aviation turbine blade is realized, and the service life of the turbine blade is prolonged.

The invention relates to a Mama-based seismic image super-resolution reconstruction method, which comprises the step of constructing a local feature extraction module, a global feature extraction module and a high-quality seismic section reconstruction module so as to solve the problem of insufficient resolution caused by high-frequency attenuation and sparse acquisition. According to the invention, a Mama model is introduced, and a novel super-resolution reconstruction network is constructed. The local feature extraction module extracts shallow feature information through a convolutional layer; the global feature extraction module obtains global features through the multi-stage residual state space block; and the high-quality seismic image reconstruction module realizes reconstruction of a high-resolution seismic image through pixel recombination and convolution operation. In addition, a Charbonnier loss function is adopted, the gradient disappearance problem is effectively avoided, and the loss function of traditional Mamba is improved. According to the method, the seismic image resolution can be improved, and meanwhile, the calculation complexity and the time cost are remarkably reduced.

The invention belongs to the technical field of time series data processing, and particularly relates to a time series data anomaly detection method based on a variable time converter. According to the method, an anomaly detection model VT-GAN is provided, the model designs a parallel generator group, and each generator extracts mode features of different time scales through expansion causal convolution; time self-attention, variable self-attention and cross attention layers are fused in the VTT architecture, and a time-space interaction relationship is explicitly modeled through a learnable gating weight; and a loss function of the VT-GAN model is adjusted and optimized in combination with temperature parameters, and data anomaly detection is carried out based on the constructed VT-GAN model. The method solves the problems that a traditional model is difficult to meet millisecond-level response requirements due to sequence calculation characteristics, the calculation complexity of a global attention mechanism in high-dimensional time series data limits the real-time performance, the false alarm rate is high, and the anomaly detection coverage rate is low.

The invention provides a lithium ion battery health state estimation method based on a TSLANet network, and belongs to the technical field of power batteries, and the method comprises the steps: obtaining multi-dimensional time sequence data of a lithium ion battery in an actual operation process, and carrying out the preprocessing of the multi-dimensional time sequence data; constructing a battery health state estimation model by using the preprocessed multi-dimensional time sequence data based on a TSLANet network and in combination with a neural network; inputting the operation condition of the current lithium ion battery system into a battery health state estimation model, and estimating the health state of the current lithium ion battery; the method is based on the TSLANet network, combines the Kalman filtering algorithm, the IC curve, the health factor and the neural network, constructs the battery health state estimation model to estimate the health state of the lithium ion battery, correctly reflects the long-term change trend of the battery performance, is excellent in noise data processing performance, and reduces the calculation complexity and the resource consumption.

Deep learning tools such as convolutional neural networks (CNNs) and transformers have spurred great advancements in computational biology. However, existing methods are constrained architecturally in context length, computational complexity, and model size. This application introduces a sub-quadratic architecture for modeling, which combines projected gated convolutions and structured state spaces to achieve local and global context with, for example, single-nucleotide resolution. These models outperform CNN-, GPT-, BERT-, and long convolution-based models in many tested genomics tasks without pre-training and with 4×-781× fewer parameters. In the proteomics domain, these models similarly outperform pretrained attention-based models, including ESM-1B and TAPE-BERT, on remote homology prediction without pre-training and while using 3,308×-23,636× fewer parameters.

The invention provides a method and a system for monitoring and predicting the health state of a storage battery based on multi-modal feature fusion, and relates to the technical field of storage battery monitoring, and the method comprises the steps: obtaining multi-modal data through periodically collecting static data of the storage battery and collecting dynamic data of the storage battery in real time; and then the method is realized through the steps of data monitoring, shared bottom layer feature extraction, SOH prediction, RUL prediction, multi-target joint optimization and the like. According to the method, multi-modal feature data generated by the storage battery can be monitored and analyzed on line, a multi-modal feature fusion and multi-target joint prediction and optimization mechanism is introduced, and SOH and RUL are predicted by adopting independent branches based on an SOH prediction model and an RUL prediction model, so that the health state of the storage battery is predicted; according to the method, automatic analysis on multi-modal mass data of the battery can be realized, the workload is reduced, the analysis efficiency is improved, the accuracy of a prediction result can be improved, and meanwhile, a constructed feature sharing mechanism can reduce the calculation complexity and improve the prediction efficiency.

The invention discloses an unmanned aerial vehicle aerial image target detection method based on PSO-DETR, and belongs to the technical field of unmanned aerial vehicle aerial image target detection. Firstly, a parallel patch perception attention feature extraction module is constructed, and an efficient multi-branch backbone network C3KCSPnet is designed by fusing a CSPDarknet53 structure; according to the network, gradient flow is improved through deep optimization, and the capturing capacity of high-level semantic information is enhanced. And secondly, an enhanced channel offset hybrid operator is provided, the dependency relationship between channels is enhanced through a channel shuffling mechanism, and cross-channel interaction of local space information is realized in combination with channel offset operation, so that the feature recovery quality and fusion efficiency in an up-sampling stage are improved, and the problem of missed detection of a shielded target is further relieved. And finally, a re-parameterization hierarchical aggregation network is designed, effective integration of shallow details and deep semantics is realized through an efficient hierarchical fusion mechanism on the premise of ensuring controllable calculation complexity, and the detection performance of the small target is further enhanced.

The invention relates to the technical field of voice processing, can be applied to business scenes of financial science and technology, medical health and the like, and discloses an audio coding and decoding method, device, equipment and medium, and the method comprises the steps: carrying out the sliding window segmentation processing of an input audio signal, and generating signal segments; processing the signal segments through an encoder containing a multi-layer self-attention mechanism to generate continuous potential representations; performing decomposition vector quantization processing on the continuous potential representation to generate a discrete code; the discrete codes are processed through a decoder comprising a multi-layer self-attention mechanism, and reconstructed signal segments are generated; and splicing the reconstructed signal segments to generate a complete audio signal. According to the method, a traditional convolution structure is replaced by a multi-layer self-attention mechanism, global time sequence dependence modeling is carried out on the audio signals after sliding window segmentation, effective compression of potential representation is realized in combination with decomposition vector quantization, and continuity and fidelity of audio reconstruction are improved on the premise that calculation complexity is not increased.

The invention relates to a mechanical arm safe obstacle avoidance control method and system based on a high-order control obstacle function, aiming at a multi-degree-of-freedom mechanical arm system with an obstacle in a task trajectory space, the high-order control obstacle function suitable for a high-relative-order system is used, and the method comprises the following steps: establishing a dynamic model of a multi-degree-of-freedom mechanical arm; designing a stable tracking controller based on the task trajectory in combination with a backstepping control method; an obstacle collision detection model is established, and safety constraints are provided for a control system through a high-order control obstacle function; and in combination with the control constraint and the optimal control problem, solving the optimal control input through quadratic programming. According to the method, the optimal task trajectory tracking can be completed on the premise that the tail end of the mechanical arm meets the safety constraint, the calculation complexity is effectively reduced, and the real-time performance is guaranteed.

The invention relates to the technical field of fluid control, and discloses a fluid control intelligent adjusting method and system for a micro-fluidic chip, and the method comprises the steps: carrying out the data collection of a self-adaptive fluid controller, a micro-fluidic channel pressure sensor array, a temperature sensor array and a flow sensor array based on the micro-fluidic chip; obtaining real-time fluid parameter monitoring data; performing feature extraction to obtain a multi-dimensional fluid feature matrix; establishing a three-dimensional fluid model of the microfluidic channel, and solving to obtain flow field distribution data; inputting the flow field distribution data into a Kriging agent model, establishing a mapping model of channel width, depth, length and fluid control performance, and calculating to obtain an optimal geometric parameter combination; and performing multi-level calculation on the fluid pressure uniformity, the temperature distribution uniformity and the flow stability to obtain a hierarchical fluid parameter control strategy. According to the invention, the problem of multi-physics field coupling is effectively solved, the calculation complexity of the control system is reduced, and the execution efficiency of the control strategy is improved.

The invention discloses a bad weather image restoration method based on a multi-modal state space model, and belongs to the technical field of computer vision. In order to solve the problem that an existing unified bad weather image restoration method has limitations in the aspects of global receptive field and computational efficiency, the unified restoration of various bad weather images is realized by designing a parallel multi-mode encoder and an adapter to generate comprehensive prompts containing degeneration semantics. Through parallel connection of a state space module and a local context sensing module of a double-attention mechanism, simultaneous capture of global long-range dependence and local detailed features is realized. Experiments show that the method is high in generalization ability, high in processing speed and low in calculation complexity, and can show good adaptive capacity on a plurality of bad weather image removal tasks.

The invention discloses a topological optimization improvement method of an initial layout for extracting mainstream characteristics under a complex heat dissipation working condition, a medium and a program product, and belongs to the technical field of intersection of thermal management and structural topological optimization, and the method is designed for a liquid cooling heat dissipation structure under the conditions of high Reynolds number turbulence and conjugate heat transfer. The method comprises the following steps: constructing a two-dimensional array type parameterized initial layout, generating high-fidelity sample data based on CFD simulation, training a neural network prediction model with an SE attention mechanism, carrying out initialization scheme screening in combination with a correction objective function under a flow field freezing hypothesis and an ETO optimization algorithm, and carrying out prediction on the flow field freezing hypothesis. And topological optimization evolution based on a variable density method is implemented on the basis of the optimization initial state. According to the method, the final effect of topological optimization can be improved while the calculation complexity is reduced, collaborative optimization of the heat dissipation efficiency and the flowing performance is achieved, and an effective means is provided for conjugate heat transfer design of the liquid cooling heat dissipation structure under the high Reynolds number turbulence complex working condition.

The invention provides an image restoration and super-resolution reconstruction system and method based on deep learning, and belongs to the technical field of digital image processing. The invention aims to solve the problems of high calculation complexity and resource consumption, limitation of long sequence processing, high training difficulty and texture scene deficiency when a multi-scale residual network based on a Transform architecture is used for image resolution conversion. The reconstruction system comprises: an image preprocessing module performing window division and video memory optimization on an input low-resolution image; the multi-layer fusion network dynamically adjusts the characteristics of the low-resolution image, captures channel information in different scenes, performs interactive fusion, performs comparison supervision, establishes an information communication channel, dynamically adjusts and optimizes parameters through negative feedback, and obtains a super-resolution image. And the loss function module maximizes the similarity of the super-resolution image and the high-resolution image in the segmentation feature space to obtain a final super-resolution image.

The invention discloses an image classification method and system based on CNN-Transform staged fusion, and the method specifically comprises the steps: inputting an input image into a lightweight CNN backbone network after preprocessing, and extracting the local features of a shallow layer, a middle layer and a deep layer; non-overlapping image blocks are divided on the middle-layer feature map, learnable position codes are embedded, and global semantic features are generated through linear attention Transform; a dynamic gating fusion module is designed, channel attention and space attention branches are combined to generate adaptive weights, and local and global features are subjected to weighted fusion; the CNN shallow detail features and Transform deep semantic features are fused through cross-level jump connection, and the classification robustness is improved. The system supports model quantization, operator fusion and edge device deployment, and the classification precision is remarkably improved while the calculation complexity is reduced.

The invention discloses a big language model knowledge graph construction method in the financial field, which performs relation extraction by introducing a deep learning algorithm, and integrates and verifies financial data from different data sources by using a field ontology and an inference engine. The problem of data integration caused by inconsistent data formats, entity redundancy and conflicts in a traditional method is solved, so that the accuracy and consistency of knowledge graph construction are greatly improved. According to the method, the problems caused by non-uniform names or data redundancy in different financial institutions, market reports and government files can be effectively eliminated, the calculation complexity is reduced, and the query efficiency is improved. And secondly, an intelligent question-answering system based on a natural language is introduced, so that the technical threshold of the user is reduced, and non-technical personnel can query and analyze financial data without mastering a complex query language.

The invention relates to a heart rate prediction and model training method and system, wearable equipment and a medium. The training method comprises the following steps: acquiring a training sample set and a corresponding label set; wherein each training sample of the training sample set comprises a photoelectric volume pulse wave sequence and a corresponding acceleration sequence, and the label set is a real heart rate; preprocessing each training sample to obtain a corresponding photoelectric volume pulse wave frequency spectrum sequence and an acceleration frequency spectrum sequence; all the photoelectric volume pulse wave frequency spectrum sequences and the acceleration frequency spectrum sequences are input into a heart rate prediction model to be trained, and corresponding predicted heart rates are obtained; and calculating the difference degree between the predicted heart rate and the corresponding real heart rate, and updating parameters of the heart rate prediction model based on the difference degree to obtain a trained heart rate prediction model. According to the method, the problems of high signal processing cost and high calculation complexity caused by the use of a complex adaptive filtering process in the prior art are solved, and the heart rate prediction accuracy can be remarkably improved.

The invention relates to the technical field of deep learning semantic segmentation, in particular to a video sequence segmentation method based on a selective scanning visual state space model, which comprises the following steps of: inputting an image sequence into an overlapped feature block division layer, and dividing the input image sequence into an image feature block sequence; inputting the image feature block sequence into a state space model encoder based on selective scanning vision to extract thickness feature sequences on different scales; inputting the coarse and fine feature sequences on different scales into a multi-layer perceptron layer for effective fusion of multi-level features to obtain a feature sequence; inputting the feature sequence into a lightweight de-wharf prediction segmentation mask based on a convolutional neural network, and performing visual visualization to generate a semantic segmentation map; in this way, the technical problems that in the prior art, the acceptance range of a traditional convolutional neural network is limited, and a self-attention-based network has extremely high calculation complexity in the aspect of constructing long-term dependence, so that the image sequence segmentation effect is poor are solved.

The invention relates to the technical field of voice processing, can be applied to service system platforms of medical health, financial science and technology, communication and the like, and discloses an audio compression and reconstruction method, device, equipment and medium. Residual vector quantization and inverse quantization are carried out to generate basic spectrum features; and compensating a high-frequency component by adopting a spectrum expansion network, predicting a time domain error in combination with a long-short-term memory network, generating a time domain residual compensation signal, and superposing the time domain residual compensation signal with the reconstructed time domain audio signal to obtain a target audio signal. Through residual vector quantization and neural network modeling, the audio reconstruction quality under low-bit-rate compression is improved, spectrum detail reservation and time domain error compensation are optimized, meanwhile, the calculation complexity is reduced, and the method is suitable for high-sampling-rate and resource-limited scenes.